Can anyone besides Nvidia make big bucks from chips?

The strange economics of the semiconductor supply chain
半导体供应链的奇怪经济学

2024年10月31日 01:06 下午

ADAM SMITH would be baffled by microelectronics. When the great economist died in 1790 James Watt’s two-cylinder steam engine passed for the height of technological sophistication. If he recognised the prefix “nano”—a fair bet for a precocious classicist proficient in dead languages by 14—it would be as a derivation of the Greek word for “dwarf”, not as a reference to the billionths of a metre in which modern semiconductors are measured. The word “billion” had entered the English language by Smith’s time but the number it denotes would have seemed unfathomable. Some 200bn transistors etched onto a few halfpennies in Nvidia’s latest Blackwell artificial-intelligence (AI) chip? Black magic, even for an enlightened Scottish rationalist.

亚当·斯密会对微电子学感到困惑。1790 年，这位伟大的经济学家去世后，詹姆斯·瓦特 (James Watt) 的两缸蒸汽机就达到了技术复杂程度的顶峰。如果他认识到前缀“nano”——对于一个 14 岁就精通已消亡语言的早熟古典主义者来说，这是一个公平的赌注——它将是希腊语“矮子”一词的派生词，而不是指十亿分之一米，其中现代半导体是可以测量的。在史密斯时代，“十亿”一词已进入英语，但它所表示的数字似乎深不可测。Nvidia 最新的 Blackwell 人工智能 (AI) 芯片中，几枚半便士的硬币上刻有大约 2000 亿个晶体管？黑魔法，即使对于开明的苏格兰理性主义者来说也是如此。

The man who observed that “the division of labour is limited by the extent of the market” would, though, be on familiar ground seeing the global chip industry. Stripped of modern jargon, the semiconductor supply chain is Smith’s insight in action. As the market’s extent has swelled, from sales of $45bn in 1979 (in today’s money) to perhaps $600bn in 2024 and, forecasters reckon, over $1trn by the early 2030s, labour has been divvied up accordingly.

不过，观察到“劳动分工受到市场范围限制”的人对全球芯片行业的看法是熟悉的。抛开现代术语，半导体供应链就是史密斯的洞察力的实际运用。随着市场规模不断扩大，销售额从 1979 年的 450 亿美元（按今天的货币计算）增加到 2024 年的 6000 亿美元，预测者估计，到 2030 年代初期将超过 1 万美元，劳动力也相应地进行了分配。

Gone are the integrated chip giants of the 1970s that, like America’s Intel back in the day, did everything themselves. This is the era of hyperspecialisation. Nvidia and other designers prepare the blueprints. “Foundries” like TSMC of Taiwan turn these into physical products by feeding silicon from one set of firms into machines assembled by another. These in turn contain gubbins from even more specialised providers, and so on.

20 世纪 70 年代的集成芯片巨头已经一去不复返了，他们就像当年的美国英特尔一样，一切都是自己做的。这是一个超专业化的时代。英伟达和其他设计师正在准备蓝图。像台湾台积电这样的“铸造厂”通过将一组公司的硅输入另一组公司组装的机器中，将这些产品转化为实物产品。这些反过来又包含来自更专业的提供商的 gubbin，等等。

What has not been split as finely is the AI windfall. Since mid-2022 the world’s 100 biggest semiconductor companies by market capitalisation have added $5.4trn in shareholder value. The 48 foundries and equipment-makers among them got just $1trn. The 32 “fabless” chipmakers, which design but do not manufacture their own processors, nabbed $4.2trn. Nvidia, whose chips are favoured by AI model-makers, captured three-quarters of that figure. (The 20 firms that continue to both design and manufacture chips, like Intel and Samsung, made barely any gains.)

人工智能带来的意外之财还没有被分割得那么细。自 2022 年中期以来，全球市值最大的 100 家半导体公司的股东价值增加了5.4 万亿美元。其中 48 家铸造厂和设备制造商仅获得 1 万亿美元。32 家设计但不制造自己的处理器的“无晶圆厂”芯片制造商获得了 4.2 万亿美元的收入。英伟达的芯片受到人工智能模型制造商的青睐，占据了这一数字的四分之三。（英特尔和三星等 20 家继续设计和制造芯片的公司几乎没有取得任何进展。）

Nvidia’s good fortune can be explained by its dizzying financial results. Its revenues nearly quadrupled between the first half of 2022 and the same period this year. It is on course to make $71bn in net profit in 2024, up from $8.4bn two years ago. Still, the foundries and toolmakers may be forgiven for feeling nettled. Nvidia’s 31 fabless peers are worth nearly as much as the foundries and equipment firms, despite bringing in half as much revenue and only two-fifths as much net profit. They also spend less than half as much on fixed assets and research and development (R&D). Moreover, the equipment-makers (though not the asset-heavy foundries) boast higher operating margins and a better return on capital than the fabless chipmakers.

英伟达的好运气可以用其令人眼花缭乱的财务业绩来解释。2022 年上半年与今年同期相比，其收入几乎翻了两番。到 2024 年，该公司净利润预计将达到 710 亿美元，高于两年前的 84 亿美元。尽管如此，铸造厂和工具制造商感到恼火也是情有可原的。Nvidia 的 31 家无晶圆厂同行的价值几乎与代工厂和设备公司相当，尽管其收入只有其一半，净利润只有五分之二。他们在固定资产和研发 (R&D) 上的支出也不到一半。此外，设备制造商（尽管不是重资产代工厂）比无晶圆厂芯片制造商拥有更高的营业利润率和更好的资本回报率。

The disparity is all the more curious in light of Smithian hyperspecialisation’s side-effect in the chip industry: its hyperconcentration everywhere except the busy fabless end. As product niches have narrowed, the chip supply chain has come to resemble a string of micro-monopolies and duopolies. TSMC and Samsung (which makes chips for clients as well as for itself) tower over the foundry business. ASML of the Netherlands is the sole purveyor of advanced lithography gear. DISCO, a Japanese company, sells 85% of the precision tools to grind and dice silicon wafers into shape. Cadence and Synopsys, two American firms, share the market for chip-design software. Advantest, from Japan, and Teradyne, from America, control post-production chip testing.

鉴于斯密式超专业化在芯片行业的副作用：除了繁忙的无晶圆厂端之外，到处都高度集中，这种差异就更令人好奇了。随着产品利基范围的缩小，芯片供应链已变得类似于一系列微垄断和双寡头垄断。台积电和三星（为客户和自己生产芯片）在代工业务中占据主导地位。荷兰 ASML 是先进光刻设备的唯一供应商。日本公司 DISCO 销售 85% 的用于研磨和切割硅片成型的精密工具。Cadence 和 Synopsys 这两家美国公司共享芯片设计软件市场。日本的Advantest和美国的Teradyne控制着生产后的芯片测试。

The niches may be small but barriers to entry are sky-high. Relative to market size, downstream firms spend huge sums on capital investments and R&D. For DISCO the annual figure is equivalent to 10% of its total addressable market; for ASML and TSMC it exceeds 20%. State-controlled Chinese challengers willing to stomach the necessary investments are being shut out of foreign markets on national-security grounds. To top it off, the incumbents can tap $150bn in handouts from Western and Japanese governments desperate to rebuild domestic chip industries.

利基市场可能很小，但进入壁垒却很高。相对于市场规模，下游企业在资本投资和研发方面投入巨额资金。对于 DISCO 来说，每年的数字相当于其潜在市场总量的 10%；ASML和台积电则超过20%。愿意接受必要投资的国家控制的中国挑战者正以国家安全为由被排除在外国市场之外。最重要的是，老牌企业可以从急于重建国内芯片产业的西方和日本政府那里获得 1500 亿美元的援助。

If all this looks like a rent-seeker’s paradise, it doesn’t feel like one, sighs a big toolmaker’s finance chief. For some micro-oligopolists that is because they sell directly to similarly domineering firms one link up the value chain. Having only a handful of buyers tempers a seller’s market power. Other companies benefit from a more diverse customer base but have preferred not to squeeze big purchasers such as TSMC, notes Chris Miller, a scholar of semiconductor history at Tufts University.

一位大型工具制造商的财务主管叹息道，如果这一切看起来像是寻租者的天堂，那么实际情况却并非如此。对于一些微型寡头来说，这是因为他们直接向价值链上的类似霸道企业销售产品。只有少数买家会削弱卖家的市场力量。塔夫茨大学半导体历史学者克里斯·米勒指出，其他公司受益于更多元化的客户群，但不愿挤压台积电等大买家。

Poolside manner

池畔方式

Quiescence is harder to sustain when demand for AI chips from deep-pocketed buyers such as Microsoft and Google lets Nvidia charge a fortune for each new model. Its gross margin is now 75%, up from 60% in 2022. Blackwells may go for $40,000 apiece, and tech giants are mulling purchases of a million or more. TSMC, the profit pool’s second-biggest bather, is expected to raise prices for cutting-edge manufacturing by up to 15%. Other firms that help turn blueprints into reality feel entitled to some of the bonanza, especially when the flow of income from China is slowing as America restricts high-tech exports to its geopolitical adversary.

当微软和谷歌等财力雄厚的买家对人工智能芯片的需求让英伟达对每个新型号收取高价时，这种平静就更难以维持。其毛利率目前为 75%，高于 2022 年的 60%。Blackwells 的售价可能为每台 4 万美元，科技巨头正在考虑购买 100 万美元或更多。利润池第二大的台积电预计将把尖端制造的价格提高最多 15%。其他帮助将蓝图变为现实的公司认为自己有权获得一些财富，特别是当美国限制对其地缘政治对手的高科技出口，来自中国的收入流正在放缓的时候。

The fabless and their suppliers could improve their profitability in tandem if all raised prices. But this might crush demand for the end product. Investors’ readiness to punish any hint of such softness, like ho-hum sales forecasts from AMD, a fabless rival of Nvidia’s, on October 29th, constrains the chipmakers’ pricing power. That may prove less of a deterrent downstream. Don’t be surprised if the reluctant oligopolists become a bit less bashful about flexing their rent-seeking muscles by the poolside. ■

如果无晶圆厂及其供应商都提高价格，他们的盈利能力就可以同步提高。但这可能会抑制对最终产品的需求。投资者准备惩罚任何此类疲软迹象，例如 10 月 29 日 Nvidia 的无晶圆厂竞争对手 AMD 做出的乏味销售预测，这限制了芯片制造商的定价能力。这可能证明对下游的威慑力较小。如果那些不情愿的寡头垄断者不再那么羞涩地在池畔展示他们的寻租肌肉，请不要感到惊讶。■